Anthropic has released a tool that embeds an invisible watermark into text generated by its Claude models, a step the company says is designed to make AI-authored content more detectable without changing how the writing reads.
The watermark is imperceptible to readers, survives copying and pasting, and will be applied to Claude-generated content globally, including through cloud service providers. Anthropic says it plans to offer third-party verification tools so publishers, universities, and others can check text against the watermark independently.
Claude models released this month or later will carry the feature by default. Anthropic says it is working to extend support to older versions of the model.
The company positioned the move as part of its obligations under the European Union AI Act, which requires greater transparency around AI-generated content.
For the publishing industry and educational institutions, both of which have struggled to establish reliable policies around synthetic writing, the watermark offers a new verification mechanism.
Anthropic was careful to set expectations around what the tool can and cannot do. Heavy editing, paraphrasing, translation, or blending Claude output with other writing may render the watermark undetectable. The company also noted that a detectable watermark does not conclusively prove Claude originally wrote the text.
The announcement places Anthropic alongside other AI developers that have explored content provenance tools, though the effectiveness of any watermarking system at scale remains an open question in the industry.
